TWO JINGJU SCHOOLS
Rafael Caro Repetto, Rong Gong, Nadine Kroher, Xavier Serra
Music Technology Group, Universitat Pompeu Fabra, Barcelona
{rafael.caro, rong.gong, nadine.kroher, xavier.serra}@upf.edu ABSTRACT
Performing schools (liupai) in jingju (also known as Pe-king or Beijing opera) are one of the most important ele-ments for the appreciation of this genre among connois-seurs. In the current paper, we study the potential of MIR techniques for supporting and enhancing musicological descriptions of the singing style of two of the most re-nowned jingju schools for the dan role-type, namely Mei and Cheng schools. To this aim, from the characteristics commonly used for describing singing style in musico-logical literature, we have selected those that can be stud-ied using standard audio features. We have selected eight recordings from our jingju music research corpus and have applied current algorithms for the measurement of the selected features. Obtained results support the de-scriptions from musicological sources in all cases but one, and also add precision to them by providing specific measurements. Besides, our methodology suggests some characteristics not accounted for in our musicological sources. Finally, we discuss the need for engaging jingju experts in our future research and applying this approach for musicological and educational purposes as a way of better validating our methodology.
1. MOTIVATION
This paper is a joint work between an ethnomusicologist (the first author) and a team of MIR researchers in the framework of the CompMusic project. In this project we exploit jingju music characteristics (and other music tra-ditions) with the aim of pushing forward the state of the art in MIR. In last ISMIR Conference (Taipei, 2014) jingju music received significant attention, with a specific tutorial and several papers published by members of our team [1-3], as well as the work by Tian et al. [4]. In the present paper though, the motivation has been to test the potential of current MIR methodologies to support and enhance qualitative and descriptive musicological anal-yses of jingju music. To this aim, we have selected one of the more relevant aspects of jingju music appreciation, which is the singing style of different performing schools;
specifically, we have focused on two of the more popular ones, Mei school and Cheng school.
The paper is structured hence in the following sec-tions. In the introduction we present the concept of jingju schools and the importance of singing style, as well as explain the purpose of the research undertaken for this paper. In the following two sections, we introduce the collection of recordings selected and the methodology proposed, and analyse the obtained results. In the discus-sion section we reflect on the challenges for expanding our research and present the direction of our future work. We conclude by summarising the musicological out-comes of the current research.
2. INTRODUCTION
Jingju is one of the genres of Chinese traditional theatre arts, arguably the most widespread and acclaimed one. Originally a folk art form, the actor traditionally was in charge of the whole creative process, from costumes and make-up, to acting, dancing, reciting, arranging the music and sometimes even writing (or improvising) the lyrics. In order to structure their performance, actors drew on a vast repertoire of predefined conventions handed down by tradition and which concerns every single aspect of this art. Characters of jingju plays are classified in acting categories or role-types, which define which set of con-ventions the actor who plays that role-type should master. The high complexity of such conventions requires the ac-tor to specialize in the performance of just one role-type during his life-time career. Along jingju history, there were some outstanding actors that excelled in the mastery of these conventions and pushed forward the artistic standards of their respective role-types or the genre as a
Figure 1. Mei Lanfang (left) and Cheng Yanqiu (right)
© Rafael Caro Repetto, Rong Gong, Nadine Kroher, Xavier Serra.
Licensed under a Creative Commons Attribution 4.0 International Li-cense (CC BY 4.0). Attribution: Rafael Caro Repetto, Rong Gong, Nadine Kroher, Xavier Serra. “Comparison of the singing style of two jingju schools”, 16th International Society for Music Information Re-trieval Conference, 2015.
whole. Some of these masters would bring their own per-sonalities to their performances and created personal styles. In a tradition that is orally transmitted, this would result in the appearance of liupai, or performing schools.1
The first half of the 20th century was the period of
ma-jor development of jingju and when the most renowned schools appeared. It saw the extraordinary development of the dan role-type, that portraying young or mid-aged female characters, but performed by male actors, due to social and political constrictions. Four of them gained the title of “four great dan actors” (si da ming dan),2 and founded their own schools. Their strong personality, the context of market competition, and even their own physi-cal condition caused schools of dan to be the ones with a greater degree of difference within one specific role-type. Among these four schools, those founded by Mei Lan-fang (1894-1961) and Cheng Yanqiu (1904-1958) (Figure 1),3 respectively named Mei and Cheng schools, are the
most widespread and followed ones today, currently per-formed in its vast majority by female actresses. These are precisely the ones we chose for our study.
Each jingju school is highly associated to a particular repertoire of plays and generally to a predominant per-formance skill. In fact, this repertoire is formed of plays arranged by the school founder to precisely showcase its mastery in that specific skill. In the case of Mei and Cheng schools, singing is the most representative and ac-claimed aspect of their art. This aspect concerns mainly two elements, the arrangement of new tunes4 and the
singing style. Among these two, the singing style is the feature that makes a performance instantly recognizable as belonging to any of these schools, and also one of the skills that performers, both professional and amateurs, put more effort to master. At the same time, it is one of the most important criteria for appreciating a performance among connoisseurs. The features that define singing style not only consist in the way voice is used in singing, but also in the very quality of the voice. Both of them should be considered not as natural personal qualities of particular actors, but as conventions that have to be trained and mastered. The resulting voice is to be under-stood hence as “an artificial voice, in the sense of
display-ing artifice, or art” [5], and followers of each school aim at mastering this voice quality as well.
Descriptions made in jingju musicology about singing style generally focus on its perceptual characteristics and
1 The translation of liupai as school can be subject to misinterpretation.
Differently to other traditions, jingju schools do not imply training in specific institutions or affiliation to specific lineages. They consist in the transmission of the performance style of individual great masters, so that the reference is always the founder of the school, and not the teach-er from whom the new actors or actresses learn.
2 Quotes from Chinese sources are given in our translation.
3 Picture from http://zh.wikipedia.org/wiki/程砚秋#/media/File:梅兰芳
与程砚秋.JPG (detail).
4 Since jingju music is created by applying pre-existing melodic
conven-tions, it is customary to use the term arrangement (bianqu) instead of composition to refer to this process.
the psychological profile of the characters conveyed through those characteristics. Wichmann [5] quotes a typ-ical description of Mei school’s singing style from
Zhongguo da baike quanshu (China Great Encyclopedia) by Hu Qiaomu, in which timbre is described as “‘sweet, fragile, clear and crisp, round, embellished and liquid.’ This timbre is considered ideal for portraying ‘natural,
graceful and poised, dignified, gentle and lovely tradi-tional women. ’” In all the musicological sources consult-ed for this paper [5-10], description of singing style al-ways includes this kind of terminology. However, since the aim of our research is add precision to musicological description, we have selected those characteristics for which audio features can be objectively computed. Table 1 shows the musicological characteristics selected and their corresponding audio features.
In the last few years there have been several studies about singing characteristics in different Chinese tradi-tional theatre genres, like jingju [11-12] and kunqu [13-14]. In these studies different role-types have been com-pared in terms of several singing characteristics by ana-lysing monophonic recordings produced by the authors. In our work, we look in depth to one particular role-type,
dan, and analyse singing characteristics with explicit ref-erence to its musicological descriptions and using com-mercial recordings. In the following section we describe the collection of recordings and explain the methodology proposed.
3. METHODOLOGY
For this study, we have selected a collection of recordings from our jingju music research corpus [1], according to two criteria: representativeness and comparability. In or-der to assure that these recordings are representative of their school, we have considered both the recording artist and the recorded aria. We have looked for artists whose
school filiation is explicitly stated in the release’s boo k-let, and arias belonging to plays for which we have liter-ary evidence (mostly from [8]) that are representative of their school. In order to maximize comparability, we have searched for plays for which musicological literature spe-cifically acknowledges a particular rendition from each of these two schools, as it is the case of Su San qijie accord-ing to [5, 8]. Since these are rare cases, due to the fact
Characteristics Audio features Pitch register Pitch histogram (1st degree)
Vibrato rate variability Vibrato rate (SD) Volume variability Loudness (SD) Brightness
Spectral centroid (mean) LTAS
Tristimulus
Timbre variability Spectral flux (mean)
Table 1. Musicological characteristics and their corre-sponding audio features considered in our study.
Figure 2. Block diagram of the methodology. Tristimulus LTAS 1st degree Rate, extent (mean, SD) Mean, SD Mean, SD Mean, SD Pitch tracks manual correction Pitch histogram computation Vibrato detection Loudness Spec. centroid Spectral flux
Audio segmentation Manual melody comp. Predominant
Harmonic model analysis-synthesis
that each school has developed its own specific reper-toires, we have also searched for arias with similar music structure. This is the case of fhc and sln, arranged in the same shengqiang and similar banshi.1 The resulting
col-lection of recordings is described in Table 2, together with the abbreviations used throughout the paper. We ar-gue that the size of this collection is appropriate for the current research since we are not performing a quantita-tive study. Instead we are using MIR methodologies for supporting qualitative descriptions.
Figure 2 describes the methodology proposed for this paper. Each of the recordings is ripped in a lossless com-pressed format with a sampling rate of 44.1 kHz. We manually identify the sections containing singing voice, for which we compute the predominant melody using the vamp plug-in version of Salamon and Gómez’s algorithm [15], setting a pitch range threshold of 100 Hz to 1000 Hz, and the voicing parameter at its maximum level, 3.0. Given that a percentage of error results, pitch tracks are manually corrected (an average of 7.16% from the com-puted frames). In order to measure pitch register, we use the methodology proposed in [16] to compute pitch his-tograms and obtain the pitch of the first degree2 from
their peaks values. The algorithm presented in [17] is used to measure vibrato rate and extent, of which we cal-culate mean and standard deviation (SD). To measure the remaining features, we separate the singing voice from the accompaniment by computing a harmonic model analysis and synthesis using the methodology presented
1Shengqiang is the musical convention in jingju that determines the
melodic skeleton; banshi refers to the metrical pattern. For more de-tailed information about these concepts, please refer to [1, 5].
2 We use “first degree” to translate gongyin. This term refers to the first
degree of the sung scale, and in jianpu notation is notated with the num-ber 1 (http://en.wikipedia.org/wiki/Numnum-bered_musical_notation). Alt-hough common functions are shared, we consciously avoid the term tonic, for its implications with tonality, absent in jingju music.
in [18], and apply to it standard algorithms for the com-putation of those features as implemented in the Essentia library [19]. For loudness we use the Loudness algorithm, normalizing the resulting mean to a factor of 0.5, so that SD is better comparable. To measure brightness, we compute mean and SD of the spectral centroid using the
Centroid algorithm. In order to better understand timbre qualities, we also compute tristimulus using the Tristimu-lus algorithm, and long-term-average spectrum (LTAS) using the implementation presented in [20]. Finally, to measure timbre variability, we compute spectral flux mean and SD using the Flux algorithm from Essentia.
In the following section, we analyse the results ob-tained for each school and relate them with their corre-sponding musicological characteristics.
School Work: Play. “Aria” (Character) Recording: MusicBrainz ID Length Artist
Mei
fhc:
Feng huan chao. “Ben ying dang
sui muqin Haojing bi nan” (Cheng Xue’e)
fhc-LYf:
a1e4b77b-88b0-4003-b688-66e39f579dc6 7:33 Li Yufu fhc-SYh:
4e3b46b2-9db7-4f52-af95-e43239a6c0e1 6:56 Shi Yihong fhc-LSs:
83d2fc7f-e1c1-4359-b417-ed9e519ecbb7 7:34
Li Shengsu ssqj:
Su San qijie. “Yu Tangchun han
bei lei mang wang qian jin” (Su
San)
ssqj-LSs:
067b8f25-888a-4a08-a495-cbc402846b10 7:15
Cheng
ssqj-CXqd:
87dbdf41-37ff-4f4a-83d4-7169d674579a 6:20 Chi Xiaoqiu sln:
Suo lin nang. “Chunqiu ting wai feng yu bao” (Xue Xiangling)
sln-CXq:
11a44af7-e29a-4c50-aa38-6139d37ca306 3:21 sln-LPh:
3dcae41a-795c-4b7d-979b-1b52aa42dd3a 3:06 Li Peihong sln-LGj:
1e705224-0b44-48aa-a0de-6386cda9d517 3:15 Liu Guijuan Table 2. Description of the recordings used in this paper. When applicable, short forms are provided.
4. ANALYSIS OF THE RESULTS
Table 3 shows a summary with the average measurement values from each of the features computed for each school.1 In this section we analyse how these results re-late to the musicological descriptions of their corre-sponding musical characteristics.
According to our musicological references, pitch reg-ister in Mei is higher than in Cheng. Since pitch range of arias for the dan role-type is consistent across plays, ap-proximately an octave and a major third, we take the pitch of the first degree as an indicator of pitch register. However, this degree is rarely sung in arias of this role-type. This is due to one singing convention, according to which female role-types shift their pitch register a fifth higher than male role-types, so that the modal center be-comes the fifth degree. To measure the pitch of the first degree hence, we compute a pitch histogram and assign a modal degree to each peak by listening to the recordings with the aid of scores. Since we observed that the peak corresponding to the sixth degree is usually the cleanest one, we take it as reference by assigning the value of 900 cents. Figure 3 shows the resulting pitch histograms for
ssqj-LSs and ssqj-CXq, compared with the equal
tempered scale. It has to be noted that in jingju there is no absolute standard pitch for tuning, but it depends on
the actor’s or actress’ needs. Notwithstanding this, a pitch is commonly assumed as reference for each role-type; for the dan role-type first degree is expected to be around E4 (329.63 Hz) [21]. Results in Table 3 show that this is the case for both schools, although first degree in Cheng is in average 6.28 Hz (33.30 cents) lower than E4, and Mei 5.78 Hz (30.09 cents) higher. Consequently, first degree in Cheng is in average 63.39 cents lower than Mei, more than a semitone. Results for all the re-cordings show that in every case first degrees from Mei are higher than those for Cheng, although the smallest difference between recordings from each school is 1.31 Hz. These results hence invite us to support the musico-logical description for pitch register.
Besides the aforementioned results, Figure 3 suggests that pitch histograms can shed light upon other aspects of singing style. Chen [22] has used histograms to study
1 Detailed results and more plots can be found in
https://github.com/jingjuschools/jingjuschoolsISMIR2015
that, as Figure 3 shows, compared with the equal tem-pered scale the fourth degree, although seldom used, is sung at a higher pitch, what is common knowledge in musicological literature. Unexpectedly, the higher octave of the first degree appears in the histograms slightly shifted higher, especially in Mei school, for which we have not found literary evidence. Besides, peak shape differences, cleaner and with lower valleys for Cheng, also suggest differences in singing style, probably re-garding vibrato and ornamentation. These observations invite us to argue that pitch histograms could be further exploited for the characterisation of singing style and ex-plore features unnoticed or not explicitly accounted for in our references.2
According to the sources consulted, Cheng “excels in
2 Differences in peaks height indicate different melodic preferences in
each school, what concerns tune arrangement, an issue not considered in this paper.
School Features
1st deg.
(Hz)
Vibrato Loudness Spectral centroid
(Hz)
Spectral flux Rate (Hz) Extent (cents)
Mean SD Mean SD SD Mean SD Mean SD
Mei 335.41 4.757 0.728 137.325 37.466 0.279 2536.739 366.968 0.121 0.063 Cheng 323.35 6.090 0.963 111.101 41.157 0.387 2136.555 451.642 0.087 0.058 Table 3. Average measurement values from each of the features computed for each school.
Figure 3. Pitch histograms for ssqj-LSs (top) and ssqj-CXq (bottom). Vertical lines show the equal tempered scale, solid red line marks the first degree, and dotted red line marks its higher octave.
using slow and fast vibratos” [9]. Consequently, the fea-ture that can better reflect this characteristic is the stand-ard deviation of vibrato rate. As can be seen in Table 3, this value is higher in Cheng than in Mei, and this is also the case for every recording. However, the difference is less than 0.3 Hz, what is barely appreciable by human ear. Variability in vibrato extent, as reflected in our re-sults, is slightly higher in Cheng than in Mei, but the dif-ference in this case is even less significant. Besides, spe-cific results from each recording are less consistent, since some instances from Mei school show higher SD in vi-brato extent than others from Cheng, and vice versa. What our results clearly show though, is that vibrato in Mei is considerably slower and wider than in Cheng, a feature that is consistent across all the recordings. Inter-estingly enough, we have not found such a remark in our musicological sources.
Results obtained for loudness variance also support musicological description, which takes volume variabil-ity as a characteristic of Cheng school compared to Mei. Results for each recording are less consistent than for other features, finding one case in Mei with higher SD than the lowest value for a recording in Cheng. These results, however, ought to be taken carefully. Firstly, they might have been affected by possibly different mix-ing levels in the production process. Secondly, bemix-ing loudness a perceptual feature, the algorithm used is an approximation to it by a simple modification of ampli-tude values, what prevent us to take them as a faithful representation of the characteristic measured.
Our musicological references agree that timbre in Mei school is brighter than in Cheng. To measure brightness, we have computed spectral centroid mean and SD. Val-ues for the mean effectively show that Mei has brighter timbre than Cheng, as an average and for every recording in each school. Given the complexity of timbre, we have also looked at LTAS, a feature that has been used in [23] to study and compare vocal tract and formant structure. Figure 4 shows the LTAS for the three performances of fhc and sln. To focus on the region with greater loud-ness, plots show the region between 200 Hz and 5000 Hz, with a logarithmic scale in the x-axis. In these plots it can be observed how frequencies over 1000 Hz are considerably higher in Mei than in Cheng, as well as the peaks with the highest loudness, contributing thus to timbre brightness. Besides this information, LTAS al-lows us to compare vocal tract between performers in each school, so that we can observe that timbre similarity is higher in Cheng than in Mei. This method hence seems promising in order to characterise individual ac-tors’ or actresses’ particularities within the overall re-quirements of the school.
Tristimulus has also been used to study and compare timbre qualities [24]. Figure 5 shows that in average both the second and the third of the three output components measured by tristumulus have higher values for Mei than for Cheng, what once more support the higher weight of higher partials in Mei. This figure also suggests a prom-ising tool for classification according to timbre quality.
Finally, results for spectral flux, computed with the aim of measuring timbre variability, show a greater value in Mei than in Cheng, a difference which is consistent across all the recordings, with special homogeneity in
Mei. Literature however remarks Cheng’s timbre vari a-bility as a defining trait of this school. It is also interest-ing to notice that the SD value for spectral centroid also shows a higher value in Cheng, although is not as con-sistent as the spectral flux values across all the record-ings. Since this descriptor was computed only for the sections of the recordings that contained singing, we re-Figure 4. LTAS for the three recordings of fhc (top)
and the three recordings of sln (bottom).
Figure 5. Scatter plot displaying values for the 2nd
and 3rd tristimulus components for each recording and
computed it setting different loudness thresholds for the transition frames in order to discard an influence of the
segmentation. Results didn’t change any tendency t o-wards a higher value in Cheng than in Mei. How to ap-proach timbre variability in jingju singing using audio analysis features remains hence an issue for future re-search, as discussed in the following section.
5. DISCUSSON
The analysis of the results presented in the previous sec-tion suggests that the methodology proposed in this paper is promising for the intended task, namely, supporting musicological description using audio analysis features. However, there are also some challenges that have to be addressed when extending this research in future work.
We aim to improve our methodology in the following two senses: automatizing most of the steps for audio analysis and improving existing algorithms to better fit jingju music characteristics. Some researchers in our team are currently working on improving the automatic segmentation method by Chen [22], adapting Ishwars’
methodology [25] to jingju music for better extraction of predominant melody, and developing new algorithms for automatic computation of the first degree. Improvement of the harmonic model analysis and synthesis as present-ed in [18] is also to be undertaken in the near future.
Arguably, the bigger challenge for the continuation of this research is gaining the engagement of jingju musi-cologists. In this paper we have aimed to show how the present approach can benefit musicological work. Yet to that aim we have consciously avoided most of the termi-nology that is more commonly used by experts when de-scribing singing style. The authors have not yet agreed on a methodology from an MIR point of view for approach-ing descriptions such as “sweet” (tian), “mellow” (run),
“fragile” (cui), “round” (yuan), or “wide” (kuan). Even when considering a characteristic like timbre variability, whose study by means of spectral flux disagrees with musicological descriptions, we wonder how much of this disagreement is due to difficulties in establishing a com-mon terminology between these two disciplines. From the field of MIR there have been recent calls for a better un-derstanding of the musical content of commonly comput-ed descriptors [26]. The authors argue that collaborative research as the one undertaken here would also encourage jingju experts to reflect on their terminology in terms of audio analysis features, and hopefully would gain com-plementary precision for those concepts.
Besides supporting musicological research, the use of audio features for qualitative analysis can be exploited for educational purposes. Jingju is a tradition that relies on oral transmission for training young actors and actresses. The key method in this training tradition consists on
“teaching by mouth and heart” (kou chuan xin shou), that is, the teacher sings and the student repeats as many times as needed for achieving an acceptable standard. Recently,
new technologies are being used as part of this process. Students use their cell phones to record their teachers, and audio and video recordings of performances are easi-ly accessible in the web. Technologies that could auto-matically evaluate the degree of similarity between the
teacher’s and the student’s performance, and moreover
offer a precise description of dissimilarities, would guide the trainee in better understanding his or her own learning process. The aim of such technologies would be perform-ing qualitative analysis of audio recordperform-ings, similar to the ones implemented in this paper. Building upon the results obtained and the methodology tested in the current work, we have started to develop such educational tools. To this goal we will require closer collaboration with jingju ex-perts. The involvement of these experts and the ac-ceptance of the educational tools by jingju trainees will also provide better evaluation methods for our research.
6. CONCLUSIONS
The current paper has studied the potential of using audio analysis features for supporting and enhancing the musi-cological description of singing style in two jingju schools for the dan role-type, namely Mei and Cheng. Our results support the description given in musicological literature for most of the characteristics analysed, and add precision to them. Pitch register in Cheng school is in av-erage 63.39 cents lower than Mei. Variability in vibrato rate is slightly higher in Cheng, what agrees with musico-logical description, but less than 0.3 Hz. Volume variabil-ity as a characteristic of Cheng school has been supported by the measure of loudness, whose SD is 38.71% higher in this school than in Mei. That this school is brighter in timbre than Cheng is supported by the mean value of its spectral centroid, 18.73% higher, but also by measure-ments in LTAS and tristimulus. Besides supporting these descriptions, our method also suggests some characteris-tics not explicitly accounted for in the sources consulted: vibrato in Mei is in average 1.33 Hz slower and 26.22 cents wider than in Cheng, pitch histograms suggest new characterisations for tuning and intonation, and LTAS looks promising for comparing vocal tracts of singers within a school. Only in the case of spectral flux, com-puted for studying timbre variability, the results do not support the musicological description, an issue that will be addressed in future research. In the light of these re-sults, we have started to extend this methodology for the development of educational tools, a project in which we hope to gain the engagement of jingju experts, who could benefit from this approach for their own research.
7. ACKNOWLEDGEMENTS
This research is funded by the European Research
Coun-cil under the European Union’s Seventh Framework Pr o-gram, as part of the CompMusic project (ERC grant agreement 267583).
8. REFERENCES
[1] R. Caro Repetto, and X. Serra: “Creating a Corpus of Jingju (Beijing Opera) Music and Possibilities for
Melodic Analysis,” ISMIR 2014, pp. 313-318, 2014. [2] A. Srinivasamurthy, R. Caro Repetto, H. Sundar,
and X. Serra: “Transcription and Recognition of
Syllable based Percussion Patterns: The Case of
Beijing Opera,” ISMIR 2014, pp. 431-436, 2014. [3] S. Zhang, R. Caro Repetto, and X. Serra: “Study of
the Similarity between Linguistic Tones and Melodic Pitch Contours in Beijing Opera Singing,”
ISMIR 2014, pp. 343-348, 2014.
[4] M. Tian, G. Fazekas, D. Black, and M. Sandler:
“Design and Evaluation of Onset Detectors Using Different Fusion Policies,” ISMIR 2014, pp. 631-636.
[5] E. Wichmann: Listening to Theatre: The Aural Dimension of Beijing Opera, University of Hawaii Press, Honolulu, 1991.
[6] H. Li 李海涓: “Jingju qingyi ‘Mei’ ‘Cheng’ er pai zai changfa shang de tong yu yi” 京剧青衣“梅”
“程”二派在唱法上的同与异 (Similarities and
differences in the singing techniques between the two schools of jingju qingyi Mei and Cheng),
Zhejiang yishu zhiye xueyuan xuebao, Vol. 11, No. 1, pp. 50-55, 2013.
[7] R. Wang 汪人立: “Mei pai changqiang yinyue de meixue pinge” 梅 派 唱 腔 音 乐 的 美 学 品 格
(Character of music aesthetics in Mei school’s
singing), Yishu bai jia, 1996, No. 1, pp. 40-47. [8] T. Wu 吴同宾, and Y. Zhou 周亚勋: Jingju zhishi
cidian 京 剧 知 识 词 典 (Dictionary of jingju knowledge), Tianjin renmin chubanshe, Tianjin, 2006.
[9] S. Yu 俞淑华: “Chuyi Chengpai de changqiang yu banzou” 刍 议 程 派 的 唱 腔 与 伴 奏 (My humble opinion about Cheng school’s singing and
instrumental accompaniment), Zuojia zazhi, 2012, No. 1, pp. 209-210.
[10] S. Yu 俞 淑 华: “Lun jingju ‘si da ming dan’ de changqiang yinse” 论京剧“四大名旦”的唱腔音 色(Discussing singing timbre in jingju’s ‘four great
dan actors’), Ming zuo xinshang, 2011, No. 33, pp. 114-115.
[11] J. Sundberg, L. Gu, Q. Huang, and P. Huang:
“Acoustical study of classical Peking Opera singing,” Journal of Voice, Vol. 26, No. 2, pp. 137-143, 2012.
[12] L. Yang, M. Tian, and E. Chew: “Vibrato
characteristics and frequency histogram envelopes in
Beijing opera singing,” 5th International Workshop on Folk Music Analysis, pp. 139-140, 2015.
[13] L. Dong, J. Sundberg, and J. Kong: “Loudness and Pitch of Kunqu Opera,” Journal of Voice, Vol. 28, No. 1, pp. 14-19, 2014.
[14] L. Dong, J. Kong, and J. Sundberg: “Long -term-average spectrum characteristics of Kunqu Opera
singers’ speaking, singing and stage speech,”
Logopedics Phoniatrics Vocology, Vol. 39, No. 2, pp. 72-80, 2014.
[15] J. Salamon, and E. Gómez: “Melody Extraction
From Polyphonic Music Signals Using Pitch
Contour Characteristics,” IEEE Transactions on Audio, Speech, and Language Processing, Vol. 20, No. 6, pp. 1759-1770, 2012.
[16] G. K. Koduri, V. Ishwar, J. Serrà, X. Serra, and H.
Murthy: “Intonation analysis of ragas in Carnatic music,” JNMR, Vol. 43, No. 1, pp. 72-93.
[17] P. Herrera, and J. Bonada: “Vibrato Extraction and
Parametrization in the Spectral Modeling Synthesis
Framework,” DAFx, 1998.
[18] X. Serra, and J. Smith: “Spectral Modeling
Synthesis: A Sound Analysis/Synthesis Based on a Deterministic plus Stochastic Decomposition,”
Computer Music Journal, Vol. 14, No. 4, pp. 12-24, 1990.
[19] D. Bogdanov, N. Wack, E. Gómez, S. Gulati, P. Herrera, O. Mayor, G. Roma, J. Salamon, J. Zapata,
and X. Serra: “ESSENTIA: an Audio Analysis
Library for Music Information Retrieval,” ISMIR 2013, pp. 423-498, 2013.
[20] T. Kinnunen, V. Hautamäki, and P. Fränti: “On the
Use of Long-Term Average Spectrum in Automatic
Speaker Recognition,” Proceedings of the 5th International Symposium on Chinese Spoken Language Precessing, pp. 559-567, 2006.
[21] B. Cao 曹宝荣: Jingju changqiang banshi jiedu: Xia
ce 京剧唱腔板式解读·下册 (Deciphering banshi
in jingju singing: Second volume), Renmin yinyue chubanshe, Beijing, 2010.
[22] K. Chen: Characterization of Pitch Intonation of Beijing Opera, Master thesis, Universitat Pompeu Fabra, Barcelona, 2013.
[23] J. Sundberg: The Science of the Singing Voice, Northern Illinois University Press, Dekalb, 1987. [24] M. Campbell, and C. Greated: The Musician’s Guide
to Acoustics, Oxford University Press, Oxford, 1987. [25] V. Ishwar: Pitch Estimation of the Predominant Vocal Melody from Heterophonic Music Audio Recordings, Master Thesis, Universitat Pompeu Fabra, Barcelona, 2014.
[26] B. Sturm: “A Simple Method to Determine if a Music Information Retrieval System is a ‘Horse’,”
IEEE Transactions on Multimedia, Vol. 16, No. 6, pp. 1636-1644, 2014.
